On automatic calibration of conceptual rainfall runoff models using optimisation techniques

Conceptual rainfall runoff (CRR) models are popular among hydrologists owing to their modest input data requirements, their simple structure and therefore their modest computational requirements. The realisation that CRR model calibration is more akin to a global optimisation problem, not a local optimisation one, was a significant development. The subsequent applications of global optimisation methods (GOM) have solved a major difficulty for estimating model parameters. Further improvement in parameter estimation may follow if constraints that help to limit the number of feasible parameter values can be developed. This study has proposed a methodology for formulating constraints related to the CRR model structure and the hydrologic data available for calibration.; For developing this methodology, the research examined the capabilities of three GOMs for calibration, namely, the Shuffled Complex Evolution (SCE) method, a genetic algorithm (GA) and a simulated annealing procedure (SA), and one local optimisation method, the downhill simplex method (DSM). The GOMs all performed better than the DSM. The SCE displayed superior accuracy and robustness for synthetic data applications, being able to find all (five) selected sets of parameter values with almost 100% accuracy. However, the GA performed better than the SCE method with real data and perhaps reflects some weakness in the SCE to find global optimal points under difficult calibration conditions. The SA was inferior to the others with both types of data applications.; The importance of selection of parameter ranges is currently given little attention in the calibration process, but even with the superior search capability of GOMs, inflated search spaces can frustrate their searches and may lead to inferior parameter estimation. Several inequalities relating model parameters with the hydrologic data were developed, which when coupled with the SCE method, significantly improved the GOM performance. This modified SCE method appeared less sensitive to the problem of parameter range specification. A method for formulating these constraints was demonstrated on synthetic data and a procedure for its application to real data was done using data from two tropical watersheds.